Research On Intelligent Detection Method Of Pavement Cracks Based On Edge Guided Network

With the development of the economy and society,the total length of national road construction continues to expand,and by the end of 2021,the total length of national highways has exceeded 5.28 million kilometers.However,due to the rapid increase in road mileage,it is difficult for road maintenance responsibility segments to timely and accurately grasp the safety situation of the responsible road sections.Therefore,efficient road maintenance and repair have become an important challenge.As a type of road disease,surface cracks on roads can reduce driving comfort and safety.Traditional road crack detection methods typically use manual inspections,which are time-consuming and laborious and are also not suitable for modern complex elevated highways and other road sections.With the decreasing cost of image acquisition equipment and the advancement of digital image processing technology,more and more road crack detections use semiautomatic or automated algorithms for image processing.However,traditional feature design requires manual processing of specific images when processing large amounts of data,which is complex and lacks generalization ability.In contrast,deep learning methods can automatically learn features and extract image information.To meet practical needs,this study conducted in-depth research on deep learning-based road crack detection methods and networks to achieve fast and accurate crack detection tasks.This study includes the following points:1)Based on the current public road crack data,a set of high-quality road crack datasets with diverse types was collected and annotated using a semi-automatic method.A data augmentation strategy for foreground-background separation was designed based on the image features of cracks and the surrounding pavement,thereby providing a large quantity and rich variety of data for subsequent training and deployment.This facilitates the investigation of model capabilities and generalization performance of different models.2)Considering that cracks are mostly linear and fine-grained targets,a bilateral segmentation structure based on the Bi Se Net V2 network was employed to enhance the model’s perception of the edges of linear objects.By incorporating an edge-guided strategy into the detail branch,the model’s ability to extract fine details,particularly the edges of linear objects,was significantly improved,leading to enhanced extraction accuracy.3)The results of the semantic branch were indexed using a quadtree and synchronized with the original images for simultaneous block processing.Detailed information extraction guided by edges was performed on the indexed blocks identified as having cracks.By optimizing the decoder part of the network’s detail branch,computational complexity was reduced,resulting in improved real-time performance.4)The integrity of the extracted cracks was repaired based on crack connectivity.Breakpoint detection,multi-feature-based breakpoint matching,and automatic calculation of gap width were employed to repair the identified gaps,ensuring the connectivity of the cracks.5)Leveraging the open-source geospatial information platform QGIS,the developed model,processing and evaluation methods for crack detection were deployed.An easy-touse crack detection plugin was developed to support disaster assessment and generate detection reports for road cracks,thereby improving the efficiency of crack detection work.